允中 发自 凹非寺 量子位 | 公众号 QbitAI 智元直接把一整套硬核装备塞进你的背包： 业界顶配机器人开发平台、海量真实数据集、以及通往机器人学术圣殿IEEE ICRA舞台的传送门 。 更有总价值超 百万美元 的奖金与资源掉落，等你来开箱！ 两大赛道，决战具身智能之巅 全球具身智能的极客玩家，是时候登场了！ 还在为缺乏 顶级硬件平台验证算法 而苦恼？还在为 仿真环境和真实产线差距太大 而挠头？还在苦恼自己的 顶尖算法只能在PPT上跑 ？ 现在，你的"天命副本"来了。 由 智元（AGIBOT） 主办的机器人领域顶级赛事—— AGIBOT WORLD CHALLENGE @ICRA 2026 ，正式向全球豪杰发出召集令！ 大赛共设置 推理-操作 和 世界模型 两大赛道，不管你擅长的是让机器人动手干活，还是让机器人动脑预判，这里都能找到你的主场。 1、Reasoning to Action（推理-操作）赛道 本赛道旨在评估模型的推理和动作执行能力，包括线上仿真和线下真机赛段。 参赛者将基于AGIBOT WORLD开源数据集，训练能够解决一系列复杂任务的模型。 本赛道聚焦于弥合Sim2Real Gap，实现从开 ...

Core Insights - The AGIBOT WORLD CHALLENGE @ICRA2026, hosted by Zhiyuan Robotics, features a total prize pool of $530,000, highlighting the significance of the event in the robotics field [3][9]. Group 1: Reasoning to Action Track - This track aims to evaluate the reasoning and action execution capabilities of models through both online simulations and real-world tasks [5]. - Participants will train models using the AGIBOT WORLD open dataset to solve complex tasks, focusing on bridging the Sim2Real gap [6]. - The competition includes various scenarios such as logistics, industrial, supermarket, dining, and home environments, with multiple tasks of varying difficulty [6]. - Zhiyuan has curated high-quality data for the competition, with each task containing hundreds of complete operation trajectories, available on hugging face and modelscope [6]. - A baseline model, ACoT-VLA, will be provided to assist participants in mastering the training, testing, and submission processes [6]. - The Genie Sim 3.0 platform, a first-of-its-kind large language model-driven open-source simulation platform, will be used for comprehensive model evaluation [6]. Group 2: World Model Track - The World Model track focuses on the core ability of embodied world models to accurately model the dynamics of physical environments based on robot actions [8]. - Participants will train video generation models using the AGIBOT WORLD dataset to generate interaction videos in ten real-world operational scenarios [8]. - The dataset consists of over 30,000 real trajectories covering diverse robot-environment interactions, including actions like grasping, placing, pushing, and pulling [9]. - The competition will utilize the EVAC model, an open-source embodied world model driven by robot actions, as the baseline model [9]. - Evaluation will be conducted using the EWMBench benchmark, assessing image quality, scene consistency, and trajectory adherence to provide reliable performance feedback [9].